| arm | n participants | n (IQR) samples per participant |
|---|---|---|
| 1 | 109 | 3 (2-4) |
| 2 | 29 | 3 (3-4) |
| 3 | 25 | 2 (2-3) |
DASSIM deep sequencing resistome analysis: bystander effect of antimicrobials on resistome
Background
DASSIM study
The DASSIM study recruited adults in Blantyre, Malawi, to three arms: 1) sepsis and antibiotic exposed (mostly ceftrixone), admitted to Queen Elizabeth Central Hospital (QECH) 2); admitted to QECH, no antibiotics 3) community members. Arms 2 and 3 were age and sex matched to arm 1. Participants had stool or rectal swab samples collected at five time points (day 0,7, 28, 90, 180) which were cultured for ESBL-Enterobacterales (ESBL-E); isolates were whole-genome sequenced. Antibiotic-exposed participants had a rapid increase in ESBL-E carriage prevalence, hospitalised participants a more modest increase, all described here.
Research questions for shotgun metagenomics
A subset of these samples (determined largely by logistics/resources) had total DNA extracted and deep sequenced (~ 11Gb). The questions we aimed to answer were
We’ve demonstrated the direct effect of antibiotic exposure on phenotypic ESBL carriage - but what are the bystander effects of antimicrobial exposure and hospitalisation on microbiome and resistome composition?
The rapid acquisition of ESBL-E in the antibiotic exposed could be due to enrichment of minority ESBL-E undetected by culture at baseline, or true hospital-associated transmission events, driven by antibiotic-induced loss of colonisation resistance. Hence:
- Is there any evidence of a role for microbiome in providing colonisation resistance against acquisition of ESBL?
- Is there any evidence for detectible ESBL genes or ESBL-E strains in samples in which no ESBL-E id detected by culture
This document describes a look at the resistome, with a focus on comparing the cultured ESBL prevalence to that detected in the deep sequence dataset, and looking at the bystander effect of antimicrobial exposure on non-ESBL genes.
Methods
AMRFinder with default settings! See sections below for methods on e.g. models as I go along.
Results
Describe number of samples
In total there are 450 samples (and 25 negative controls). Table 1 shows number of participants and number of samples per participant.
Describe resistance genes present
Table 2 and Figure 1 shows the prevalence of AMRFinder-defined AMR gene class, stratified by arm in the latter case; beta-lactam, trimethorpim, tetracyycline, sulfonamide and lincosamide/macrolide/streptogramin class genes all have prevalence > 90%. Note also that there are two samples in which plasmid mediated colistin resistance genes are identified: these are mcr10 and mcr10.1, both in admission samples of participants without prior hospitalisation in the last 3 months.
Figure 1 suggests that there are some differences in prevalence in quiniolone and aminoglycoside/quinolone AMR genes between arms (higher in arm 1); plus possibly phenicol, macrolide - these differences are explored below.
Figure 3 shows a simple plot of number of samples in which a given gene was identfied; in the interests of space only those genes present in 5 or more samples are shown.
| AMR gene class | Prevalence (95% CI) |
|---|---|
| beta-lactam | 95% (92-97%) |
| tetracycline | 95% (92-97%) |
| trimethoprim | 95% (92-97%) |
| sulfonamide | 94% (92-96%) |
| aminoglycoside | 92% (89-95%) |
| lincosamide/macrolide/streptogramin | 92% (89-95%) |
| macrolide | 77% (73-81%) |
| macrolide/streptogramin | 61% (56-65%) |
| phenicol | 60% (55-64%) |
| nitroimidazole | 38% (33-42%) |
| fosfomycin | 30% (26-34%) |
| glycopeptide | 30% (26-34%) |
| streptothricin | 25% (21-29%) |
| lincosamide | 22% (18-26%) |
| quinolone | 21% (17-25%) |
| aminoglycoside/quinolone | 20% (17-24%) |
| phenicol/quinolone | 17% (14-21%) |
| lincosamide/macrolide | 14% (11-17%) |
| rifamycin | 7% (5-10%) |
| lincosamide/streptogramin | 4% (3-7%) |
| bleomycin | 1% (0-3%) |
| colistin | 0% (0-2%) |
| fusidic acid | 0% (0-1%) |
| pleuromutilin | 0% (0-1%) |
| streptogramin | 0% (0-1%) |
Closer look at beta lactamases
Focus now on beta lactamases; Figure 4 shows the number of genes present stratified by AMRFinder defined beta-lactam subclass. A number of these (highlighted in red) are bacteroides-specific and so less clinically relevant - they will be excluded from some downstream analyses.
Quite surprising perhaps is the number of carbapenemases; 7% of participants in arm 1 and 2 have a carbapenemase gene identified at least once - see Table 3.
Figure 5 shows prevalence of beta lactamase by participant stratified by visit and arm. Arm 1 here are antbiotic exposed by D7 (usually ceftriaxone) which accounts for the rise in ceftriaxone resistance determinants and phenotypic ESBL. See below for assessment of antimicrobial exposure on other resistance determinants. Important to note here also that the prevalence of phenotypic ESBL is mostly more than detection of cephalosporin resistance gene at most timepoints- in fact concordance between the tests is not great - Table 4 expresses this as sensitivity of gene presence for an ESBL phenotype reference standard. Some of this I think could be improved by curating the AMRFinder defined gene classes to restrict to ESBL; it seems to include (for example) OXA-1?
| class | Arm 1 | Arm 2 | Arm 3 |
|---|---|---|---|
| beta-lactam | 107/109 (98% [94-100%]) | 29/29 (100% [88-100%]) | 18/25 (72% [51-88%]) |
| carbapenem | 8/109 (7% [3-14%]) | 2/29 (7% [1-23%]) | 0/25 (0% [0-14%]) |
| cephalosporin | 84/109 (77% [68-85%]) | 18/29 (62% [42-79%]) | 9/25 (36% [18-57%]) |
| ESBL | absent | present | Total | summary |
|---|---|---|---|---|
| Negative | 82 | 35 | 117 | Specificity = 70% |
| Positive | 74 | 145 | 219 | Sensitivity = 66% |
| Total | 156 | 180 | 336 |
Bystander effect of antibiotic exposure
Look now at prevalence of all antibiotic classes stratified by arm and visit. The aim here is to assess the effect of antimicrobial exposure on AMR genes other than beta lactamases. To inform this, the antibiotic exposures in arm 1 are shown in Figure 6 (taken from DASSIM publication at https://doi.org/10.1093/cid/ciab710). See below for a more sophisticated attempt to link antimicrobial exposure to AMR gene acquisition. The prevalence of AMRFinder-defined AMR gene subclass by arm and visit is shown in Figure 7 - note that this is subclass rather than class (as is shown in Figure 1, which is why the labels are not entirely clear )
This suggests that several gene subclasses increase in response to antimicrobial exposure (i.e. they increase at visit 1 then fall off):
- amikacin/kanamycin/quinolone/tobramycin
- aminoglycoside
- azithromycin/erythomycin/spiramycin/telithromycin
- azithromycin/erythromycin/streptogramin b/tylosin
- gentamicin
Some other genes increase but it is a bit less clear cut (overlapping confidence intervals):
- clindamycin/erythromycin/streptogramin b
- rifamycin
And vancomycin resistance genes seem to decrease at visit 1 - though this also is not clear cut. The actual genes within these classes are shown in Figure 8.
Some interesting bits to pull out here - there was no macrolide exposure in the cohort but the macrolide-lincosaimde-strepotgramin (MLS) resistance gene mphA is very relevant in this setting, as it confers resistance to azithromycin in Salmonellae (and other enterics). Also no aminoglycoside exposure but lots of increase in aminoglycoside genes. I quantify these effects with a modelling approach below.
Models of effect of antimicrobial exposure on resistome
Here I’ve put some simple models of antimicrobial exposure together. Start with a simple radfnom intercept model per participant. In this model, the outcome is binary - presence or absence of any gene of the given subclass. The predictor variables are the same for each model:
- ceftriaxone exposure
- hospitalisation
- cotrimoxazole exposure
- ciprofloxacin exposure
- amoxicillin exposure
These are selected as they are the exposures that at least 10% of the cohort have experienced. Time-dependence is incorporated by allowing the effect of each exposure to decay in an exponential fashion, defined in the model by a time constant tau - the rate of decay is fitted by the model, and is defined to be the same for all exposures.
Random intercept model
Simplest is a random intercept model. Consider \(n\) measurements of a binary outcome for \(N\) participants; each may have a different number of measurements. We can construct a logistic regression model where the measurements \(y_{i}, i = 1,2 ... n\) are given by:
\[ y_{i} \sim \text{Bernoulli}(p_{i}) \] and
\[ p_{i} = \text{logit}(\alpha + \sum_{j=1}^{j=n}{ x_{ij} \beta_{j}} + \sum_{k=1}^{k=N}{z_{ik}\gamma_{k}}) \]
Where \(\alpha\) is the model intercept; \(x_{ij}\) is the value of the \(j\)th covariate for observation \(i\); \(\gamma_{k}\) is the value of the random effect for patient \(k\) and \(z_{ik}\) is the width \(N\), height \(n\) design matrix that encodes which sample belongs to which patient, and the random intercept is fixed for each patient, no matter how close together or far apart in time the samples are, and drawn from a normal distribution:
\[ \gamma_{i} \sim \text{Normal}(0, \sigma) \]
These models are not a great fit to the data (not shown here).
Priors
The priors on \(\boldsymbol{\beta}\) and \(\beta_{0}\) are similar to previous (i.e. can be student t distribution or similar) and priors on \(\alpha\) and \(\sigma\) (which are essentially scale parameters) can be similar. However to prior on \(l\) needs some thought or it can easily mess up the exploration of the posterior either if the prior puts to much weight on length parameters that imply correlation on a scale less than or more than the separation of the measurements. Following Stan user manual and some experimentation, an inverse gamma prior with parameters (2,2) seems pretty good.
Final models
So the final model becomes
\[ \boldsymbol{y} \sim \text{Bernoulli}(\boldsymbol{p}) \] Where the length \(n\) vector \(\boldsymbol{y}\) is the vector of observations and \[ \boldsymbol{p} = \text{logit}(\beta_{0} + \boldsymbol{X}\boldsymbol{\beta} + \boldsymbol{f}) \] Where \[ \boldsymbol{f} = \boldsymbol{L\eta} \] \[ \boldsymbol{\eta} \sim \text{Normal}(\boldsymbol{0}, \boldsymbol{I}) \] \[ \boldsymbol{L}^{T}\boldsymbol{L} = \boldsymbol{\Sigma} \]
Where the covariance matrix \(\boldsymbol{\Sigma}\) for two samples \(i\) and \(j\) seperated by time \(t_{ij}\) is defined by
\[ \Sigma_{ij} = \alpha^{2} \text{exp}(\frac{-t^{2}_{ij}}{2l^{2}}) + \delta_{ij}\sigma^2 \]
Where \(\delta_{ij}\) = 1 for \(i=j\) and \(0\) otherwise.
\(\boldsymbol{X}\) is the matrix of covariates for a given sample (where 1 is present, 0 absent) and the value of any entry of this matrix encoding antibiotic exposure is 1 whilst exposure is ongoing and is allowed to decay a time \(t\) following cessation of exposure as:
\[ \text{exp}(\frac{-t}{\tau}) \]
With priors \[\beta_{0}, \boldsymbol{\beta}, \sigma, \alpha, \tau \sim \text{studentT}(3\text{df}, 0, 3)\] \[l \sim \text{InvGamma}(2,2 )\]
Where the time of sample collection variable is mean centred and scaled and time since antibiotic exposure is scaled by the SD of the time variable.In practice, we can fit each participant separately to avoid making the gigantic covariance matrix.
Results
Based on the model diagnostics (traceplots and estimated bayesian fraction of miossing information EBFMI) this meodel is a better fit to the data from previous. Parameter estimates from the models are shown in Figure 9. The conclusions are largely unchanged from the simpler random intercept models.
These largely come up with the same conclusions as eyeballing the plots above. Main takeaways:
cephalosporin exposure is associated with cephalosporin resistance but also 1) lots of MLS (macrolide-lincosaimde-strepotgramin) group resistance determinants and 2) aminoglycoside resisatnce determinants. Also looks to be associated with rifamycin resistance genes.
Cotrimoxazole has wide confidence intervals - some MLS gene association, but effect seems less clear than other drugs.
Ciprofloxacin - wide confidence intervals but some clear effects - quinolone/aminoglycoside association (which I guess would be expected) but some MLS genes - and cephalosporina and chloramphenicol (though CIs cross 0). Interestingly though it seem sto be associated with a reduction in lincosamide and macrolide resistance genes.
Amoxicillin - again wide CIs make it hard to draw any conclusions I think.
Conclusions
Conclsuions here with the more complex models are pretty much unchanged form the simpler random effects models.
I think there are a few interesting conclusions from this:
There are some interesting genes in there!
- mcr plasmid-mediated colistin resistance genes - I don’t recall these being detected in Malawi before (though have not looked specifically for publications)?
- Way more carbapenemases than we found in WGS - mostly OXA though also some I’m not very familiar with and would need to check out how clinically relevant they are.
ESBL-E prevalence by culture is higher at most time points then ESBL gene detection - i.e culture is more sensitive to ESBL-E detection (probably) than shotgun metagenomics. I mean, this is not surprising but these data are probably not going to help us answer the question as to whether ESBL-E that appear in culture following antibiotic exposure are actually present at low level and enriched for by antibiotic exposure.
Cephalosporin exposure is associated with increase in cephalosporin resistance genes - I mean, we knew this from culture but it is reassuring that this is true here.
Cephalosporin exposure has a pretty clear bystander effect on other resistance genes; it is is associated with an increase in:
- MLS (macrolide-lincosaimde-strepotgramin) resistance genes
- aminoglycoside resistance genes
- rifamycin resistance genes
Ciprofloxacin also has an effect on quinolone/quinolone-aminoglycoside genes and a bystander effect on some MLS genes; and seems to be associated with a reduction in lincosamide and macrolide genes. Though the effects here are less clear cut because quinolone exposure is much less common.
Discussion/next steps
I think the mechanism of the resistome bystander effect can be defined - is this co-selection of resistance determinants carried by ESBL Gram negatives (plausible with aminoglycoside genes) or antibiotic exposure allowing proliferation of other species (Perhaps for MLS genes - ? enterococcus given cephalosporin resistance). We could do this by linking AMR gene and taxonomy either by the binned assemblies or just looking at correlation of microbiome composition and AMR gene presence (or, probably, both).
It would be great to incorporate Viv’s microbiome composition analysis into this both just as a description with a closer link to the antibiotic exposure metadata (to pin down antibiotic effect on microbiome composition) and to link to resistome as above.
Also would like to look at the role of microbiome composition in providing colonisation resistance to ESBL and whether “acquired” ESBL-E strains following antibiotic exposure can be detected in initial samples - I guess these are for discussion.
Also for discussion - Ed’s work on the salmonella plasmid and how it fits in with this.